Zhang et al. Clinical Epigenetics (2017) 9:9 DOI 10.1186/s13148-017-0315-9

SHORTREPORT Open Access Individual CpG sites that are associated with age and life expectancy become hypomethylated upon aging Yan Zhang1, Jan Hapala2,3, Hermann Brenner1,4 and Wolfgang Wagner2,3*

Abstract Background: There is a growing interest in simple molecular biomarkers for biological aging. Age-associated DNA methylation (DNAm) changes at specific CG dinucleotides can be combined into epigenetic age predictors to estimate chronological age—and the deviation of chronological and predicted age (Δage) seems to be associated with all-cause mortality. In this study, we have further validated this association and analyzed whether or not individual age-associated CG-dinucleotides (CpGs) are related to life expectancy. Findings: In the German ESTHER cohort, we used 864 DNAm profiles of blood samples as the discovery set and 1000 DNAm profiles as the validation set to predict chronological age with three previously reported age predictors—based on 99, 71, or 353 age-associated CpGs. Several of these individual CpGs were significantly associated with life expectancy, and for some of these CpGs, this was even reproducible in the independent datasets. Notably, those CpGs that revealed significant association with life expectancy were overall rather hypomethylated upon aging. Conclusion: Individual age-associated CpGs may provide biomarkers for all-cause mortality—but confounding factors need to be critically taken into consideration, and alternative methods, which facilitate more quantitative measurements at individual CpGs, might be advantageous. Our data suggest that particularly specific CpGs that become hypomethylated upon aging are indicative of biological aging. Keywords: DNA methylation, Epigenetic, Aging, Mortality, Life expectancy, Predictor

Findings reproducible DNA methylation (DNAm) changes at Biomarkersforagingmayallowfortestingofinter- specific sites in the genome [4–8]. About 60% of ventions to extend lifespan or to increase the odds of these age-associated CG dinucleotides—so called staying healthy. Ideally, such biomarkers should rather “CpG sites”—become hypomethylated upon aging, reflect “biological age” than “chronological age,” and whereas about 40% become hypermethylated [9]. Age- they should not be skewed by predisposition to spe- associated hypermethylation is rather enriched close cific diseases [1]. Advances in molecular biology, gen- to CG islands (CGIs), whereas hypomethylation rather etics, and epigenetics have fueled the hope for simple occurs outside of CGIs [9–12]. Furthermore, particu- and reliable biomarkers for biological age [2, 3]. larly DNAm at CpGs with age-associated hypermethy- Withinthelastfiveyears,amultitudeofstudies lation seem to be coherently modified in cancer [13], demonstrated that aging is associated with highly indicating that de novo DNAm and demethylation may be regulated by different mechanisms. It is yet unclear how these DNAm patterns are regulated, and * Correspondence: [email protected] 2Helmholtz-Institute for Biomedical Engineering, Stem Cell Biology and if they are functionally relevant or rather reflect other Cellular Engineering, University Hospital of the RWTH Aachen, Pauwelsstrasse means of chromatin conformation—either way, they 20, 52074 Aachen, Germany provide powerful biomarkers. 3Institute for Biomedical Engineering—Cell Biology, University Hospital of the RWTH Aachen, Pauwelsstrasse 30, 52074 Aachen, Germany Several age-associated DNAm changes are acquired Full list of author information is available at the end of the article linearly over time and hence facilitate estimation of

© The Author(s). 2017 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Zhang et al. Clinical Epigenetics (2017) 9:9 Page 2 of 6

chronological age—either based on individual CpGs [14] al. [16], and Weidner et al. [17] in the discovery and or by integration of multiple CpGs into age predictors validation sets, as well as in the overall population [5, 6, 12]. Particularly, the epigenetic clock described by (Table 1). Overall, all three models revealed good cor- Horvath [15], consisting of 353 age-associated CpGs, has relation with chronological age, albeit the correlation been shown to facilitate precise age estimations across was slightly lower for the Weidner model (Fig. 1a, b). multiple tissues. Other frequently used age predictors On the other hand, epigenetic age predictions of the for blood samples have been introduced by Hannum and Hannum predictor were on average overestimated by coworkers (71 CpGs) [16] and Weidner et al. (99 CpGs) 5.5 years in the discovery set and 6.5 years in the val- [17, 18]. Notably, the difference between chronological idation set (Fig. 1c, d). Hence, the mean average devi- age and predicted age—referred to as Δage—seems to be ation (MAD) of predicted and chronological age was related to the parameters of biological aging: Marioni higher for the Hannum predictor in the discovery and and coworkers have demonstrated that Δage (per 5 years) validation set than for the other two predictors was associated with a 21% higher mortality risk in the (Table 1). Such shifts do not affect inter-quartile com- “Hannum predictor” (95% CI 1.14–1.29) and with a 11% parison, Cox regression analysis, or hazard ratios, higher mortality risk with the “Horvath predictor” (95% which are usually described in the literature. However, CI 1.05–1.18), if adjusted for chronological age and gen- they have impact on Δage and should therefore be der [19]. Similar findings were reproduced by other taken into consideration if Δage is addressed for indi- study groups on other datasets [18, 20, 21]. Furthermore, vidual patients or for direct comparison of different epigenetic age predictions are lower in women and in datasets. It is conceivable that the higher MAD in one semi-supercentenarians [22], whereas accelerated epi- or the other epigenetic age predictor is due to preva- genetic age was associated with obesity [23] and with lence of specific diseases. “Healthy subjects” are diffi- lower abilities in physical and mental fitness [24]—sug- cult to define, and therefore, we have exemplarily gesting that age-associated DNAm patterns may be indi- excluded participants with prevalent diabetes, cardio- cative of biological aging. vascular disease, and a history of cancer at baseline In this study, we aimed for a better understanding of how (discovery panel: 180, 189, and 75, respectively; valid- epigenetic age predictions are associated with life expect- ation set: 162, 182, and 66, respectively). Removal of ancy in the ESTHER study cohort, a large population-based these participants resulted in a very similar distribu- epidemiological study conducted in the German State of tion of age predictions, indicating that general offset Saarland. To estimate reproducibility of results, we sepa- oftheagepredictorswasnotduetothesechronicdis- rated the DNAm profiles (analyzed by HumanMethylation eases (Additional file 1: Figure S1). 450 BeadChips) into a discovery set of 864 samples and a Previous studies have demonstrated that Δage of the validation set of 1000 samples (further information is pro- Hannum and Horvath predictors are associated with vided in the Additional file 1). We were particularly inter- life expectancy in DNAm profiles of the ESTHER ested whether there are individual CpGs that reveal higher study[20].Here,wehaveanalyzedifΔage of the association with life expectancy than others. Weidner model would also be associated with all- cause mortality. When the results were adjusted for Comparison of different multi-CpG age predictors age, sex, batch, and leucocyte distribution, there was a Initially, we compared epigenetic age predictions of clear tendency in the discovery and validation sets, but the three aging models by Horvath [15], Hannum et the results did not reach statistical significance (P =

Table 1 Correlation of age predictions with chronological age Weidner99 CpGs (61 hypo- Hannum71 CpGs (31 hypo- Horvath 353 CpGs (186 hypo- and 38 hypermethylated) and 40 hypermethylated) and 167 hypermethylated) Discovery set (n = 864) Correlation with age (Spearman) 0.705 0.809 0.761 Mean average deviation (years) 4.76 5.82 4.19 Validation set (n = 1000) Correlation with age (Spearman) 0.712 0.774 0.750 Mean average deviation (years) 4.78 7.00 3.95 Overall (n = 1864) Correlation with age (Spearman) 0.705 0.787 0.753 Mean average deviation (years) 4.75 6.45 4.06 Zhang et al. Clinical Epigenetics (2017) 9:9 Page 3 of 6

Fig. 1 Correlation of predicted age with chronological age. Epigenetic age predictions based on the 99 CpGs of the Weidner predictor [17] were plotted against chronological age for a 864 DNAm profiles of the discovery set and b 1000 DNAm profiles of the validation set of the ESTHER cohort. The distribution of chronological age and predicted age with the three aging models described by Weidner et al. [17], Hannum et al. [16], and Horvath [15] is demonstrated c for the discovery set and d for the validation set. Age predictions by the Hannum predictor were overall overestimated by 5.5 and 6.5 years, respectively

0.058 and P = 0.095, respectively). When we combined These results support the notion that the association of the discovery and validation sets to increase statistical Δage with all-cause mortality may vary between different power, the results reached the significance (P =0.041)and aging models and cohorts—but it is overall consistent if the hazard ratios were slightly lower than in the other two using age predictors that comprise multiple CpGs. predictors (HR = 1.087; 95% CI 1.003–1.178; Additional file 1: Table S1). In our previous work, we analyzed the Individual CpGs are associated with life expectancy data of the Lothian Birth Cohort 1921 (LBC1921), a study We have previously analyzed if individual age-associated from the Lothian region (Edinburgh and its surrounding CpGs are associated with life expectancy in the Lothian areas of Scotland) with participants born in 1921 and Birth Cohorts 1921 and 1936 [18]. The only one CpG site analyzed at about the age of 79 [18, 25]: in this dataset a that reached statistical significance in both datasets after 5-year higher age prediction by the Weidner model was multiple correction and adjustment for age and gender was associated with 11% greater mortality risk (P = 0.0003; cg05228408, which is associated with the gene for the 95% CI 1.04, 1.19; after adjustment for age and gender). chloride transport protein 6 (CLCN6; LBC1921 [HR = 1.16; Zhang et al. Clinical Epigenetics (2017) 9:9 Page 4 of 6

95% CI 1.06–1.26; P = 0.00072]; LBC1936 [HR = 1.26; 95% selected to work together, rather than individually. Fur- CI 1.12–1.42; P = 0.00013]). This genomic region is of spe- thermore, the Horvath predictor was trained on multiple cific interest because single-nucleotide polymorphisms tissues rather than blood samples as in the Hannum and identified in its vicinity were found to be associated with Weidner predictors. blood pressure and hypertension [26–28]. Therefore, we To our surprise, almost all of the CpGs that are associ- have now trained a model for the ESTHER discovery group ated with life expectancy in either of the two datasets based on the beta values of cg05228408. Upon the adjust- were hypomethylated upon aging (Fig. 2b, c). In the dis- ment for chronological age, gender, batch, and leucocyte covery set there was a significant enrichment of hypo- distribution, this model revealed significant association with methylated CpG sites (hypergeometric distribution) for − all-cause mortality in the discovery (P = 0.0011) and in the the Weidner (P = 3.3 × 10 6) and the Hannum (P = overall population (P = 0.0148; Additional file 1: Table S2). 0.0007) predictor. Furthermore, all significant CpGs in Subsequently, we tested the association with life ex- the overlap of the discovery and the validation set were pectancy for all individual CpGs of the three age predic- hypomethylated (Additional file 1: Table S6). tors: for 99 CpGs of the Weidner predictor (Additional We revisited the previously published data on association file 1: Table S3), for 71 CpGs of the Hannum predictor of these CpGs in the Lothian Birth Cohort 1921 [18]. A big (Additional file 1: Table S4), and for the 353 CpGs of the advantage in this cohort is that it comprises donors of a de- Horvath predictor (Additional file 1: Table S5). In the fined age range (about 79 years)—and hence, a different discovery set, 27 (of 99 CpGs), 11 (of 71 CpGs), and 3 slope in the comparison of predicted and chronological CpGs (of 353 CpGs) reached statistical significance ages would hardly affect the association with life expect- (FDR < 0.05). In the validation set, with a lower number ancy. Only four CpGs of the Weidner predictor reached of death cases, it was only 11, 7, and 3 CpGs, respect- statistical significance in LBC1921 (adjusted P value <0.05), ively (Fig. 2a). Albeit the reproducibility between the two and all of them were also significant in the ESTHER discov- datasets was not very high, there was a significant associ- ery set: cg05228408 (CLCN6), cg12554573 (PARP3), ation for the 99 CpGs of the Weidner predictor (hyper- cg25268718 (PSME1), and cg03224418 (SAMD10)—- geometric distribution: P value = 0.0072) and for the furthermore, all of them become hypomethylated upon Horvath predictor (P value = 0.025; Additional file 1: aging (Additional file 1: Figure S2A). However, for the Table S6). The CpGs that were overlapping associated CpGs of the Hannum predictor, the reproducibility be- with life expectancy in both datasets were cg05294455 tween the LBC1921 and the ESTHER cohorts was low. (MYL4), cg08598221 (SNTB1), cg09462576 (MRPL55), In general, CpGs that revealed significant association cg15804973 (MAP3K5), cg20654468 (LPXN), cg25268718 with life expectancy in LBC1921 and LBC1936 were ra- (PSME1), cg26581729 (NPDC1), and cg02867102 (no ther hypomethylated, but these results did not reach gene). Please note that the number of individual CpGs statistical significance (Additional file 1: Figure S2B, C). that reached statistical significance in the three predictors is not a quality measure for these age predictors. The Conclusions CpGs of the Hannum and Horvath predictors were se- Our explorative study further supports the notion that lected by Elastic Net algorithms—they were therefore specific age-associated CpGs can be indicative of life

Fig. 2 CpGs that correlate with all-cause mortality are hypomethylated upon aging. a For all individual CpGs of the three age predictors (Weidner

et al., 99 CpGs; Hannum et al., 71 CpGs; and Horvath, 353 CpGs), the association of Δage with all-cause mortality was estimated. The P values in the discovery and validation sets of the ESTHER cohort demonstrate moderate reproducibility between the two independent datasets. b, c Subse- quently, we analyzed the Spearman correlation of these CpGs with chronological age. CpGs with significant association with all-cause mortality were overall hypomethylated upon aging (in the discovery set (b) and in the validation set (c)). The lines indicate a FDR significance level of 0.05 Zhang et al. Clinical Epigenetics (2017) 9:9 Page 5 of 6

expectancy, but the reproducibility in independent publically available repositories. Individual data access may be granted within cohorts is overall not very high. Furthermore, we a framework of scientific cooperation. demonstrate that significant association with all-cause Authors’ contributions mortality is particularly observed in CpGs that become YZ, HB, and WW conceived the study. YZ performed bioinformatics analysis. hypomethylated upon aging. It is therefore conceivable JH and WW performed cross validations. WW wrote the first draft of the that a combination of such specific age-associated manuscript. All authors read and approved the final manuscript. CpGs gives rise to alternative epigenetic age predictors Authors’ information that better reflect the association of Δage with all-cause Not applicable mortality—and may hence be a better biomarker for Competing interests biological aging. WW is involved in the company Cygenia GmbH that may provide service for There are however limitations that need to be critically epigenetic age predictions to other scientists (www.cygenia.com). The taken into consideration: (1) only blood samples have been authors’ declare that they have no competing interests. considered for this analysis, and it remains to be demon- Consent for publication strated if the findings hold also true for cells from other tis- Not applicable sues; (2) the association of life expectancy with CpGs that become hypomethylated upon aging was only addressed on Ethics approval and consent to participate The ESTHER study was approved by the ethics committees of the University elderly people, whereas biomarkers for biological aging may of Heidelberg and of the state medical board of Saarland, Germany. All rather be desired for young humans who had not yet devel- participants provided written informed consent. oped age-related diseases [29]; (3) Δ of epigenetic age age Author details predictions may have systematic offsets, and hence, it re- 1Division of Clinical Epidemiology and Aging Research, German Cancer mains a challenge to entirely rule out that the results are Research Center (DKFZ), Im Neuenheimer Feld 581/TP4, 69120 Heidelberg, 2 impacted by chronological age; (4) the beta values of Illu- Germany. Helmholtz-Institute for Biomedical Engineering, Stem Cell Biology and Cellular Engineering, University Hospital of the RWTH Aachen, mina BeadChip correlate with the absolute level of DNAm, Pauwelsstrasse 20, 52074 Aachen, Germany. 3Institute for Biomedical but the precision is not always high [30]. Particularly, for Engineering—Cell Biology, University Hospital of the RWTH Aachen, 4 age predictors based on individual CpGs, it therefore ap- Pauwelsstrasse 30, 52074 Aachen, Germany. Network Aging Research (NAR), University of Heidelberg, Bergheimer Strasse 20, 69120 Heidelberg, Germany. pears to be advantageous to train model on data that was generatedbymorequantitativemethods—such as pyrose- Received: 1 December 2016 Accepted: 19 January 2017 quencing, MassARRAY, bisulfite deep sequencing, or digital PCR [18]; and (5) last but not least, the association with all- References cause mortality is only one aspect of biological aging, and it 1. Baker GT, Sprott RL. Biomarkers of aging. Exp Gerontol. 1988;23:223–39. will be important to better understand the association with 2. Burkle A, Moreno-Villanueva M, Bernhard J, Blasco M, Zondag G, et al. MARK-AGE biomarkers of ageing. Mech Ageing Dev. 2015;151:2–12. other molecular parameters, such as telomere length, or 3. Belsky DW, Moffitt TE, Cohen AA, Corcoran D, Horvath S, et al. Telomere, functional measures, such as physical strength, cognitive epigenetic clock, and biomarker-composite quantifications of biological decline, and other signs of aging [3]. aging: do they measure the same thing? bioRxiv 2016;doi: http://dx.doi.org/ 10.1101/071373. 4. Bork S, Pfister S, Witt H, Horn P, Korn B, Ho AD, Wagner W. DNA Additional file methylation pattern changes upon long-term culture and aging of human mesenchymal stromal cells. Aging Cell. 2010;9:54–63. 5. Koch CM, Wagner W. Epigenetic-aging-signature to determine age in Additional file 1: This file contains additional details on the methods, different tissues. Aging (Albany NY). 2011;3:1018–27. Additional file 1: Figures S1–S2, and Additional file 1: Tables S1–S6. (PDF 1054 kb) 6. Bocklandt S, Lin W, Sehl ME, Sanchez FJ, Sinsheimer JS, Horvath S, Vilain E. Epigenetic predictor of age. PLoS One. 2011;6:e14821. 7. Teschendorff AE, Menon U, Gentry-Maharaj A, Ramus SJ, Weisenberger DJ, Acknowledgements et al. Age-dependent DNA methylation of genes that are suppressed in Not applicable stem cells is a hallmark of cancer. Genome Res. 2010;20:440–6. 8. Rakyan VK, Down TA, Maslau S, Andrew T, Yang TP, et al. Human aging- associated DNA hypermethylation occurs preferentially at bivalent Funding chromatin domains. Genome Res. 2010;20:434–9. The ESTHER study was supported by the Baden-Württemberg State Ministry of 9. Johansson A, Enroth S, Gyllensten U. Continuous aging of the human DNA Science, Research, and Arts (Stuttgart, Germany), the Federal Ministry of Educa- methylome throughout the human lifespan. PLoS One. 2013;8:e67378. tion and Research (Berlin, Germany), and the Federal Ministry of Family Affairs, 10. McClay JL, Aberg KA, Clark SL, Nerella S, Kumar G, et al. A methylome-wide Senior Citizens, Women, and Youth (Berlin, Germany). The sponsors had no role study of aging using massively parallel sequencing of the methyl-CpG- in the study design, in the collection, analysis and interpretation of data, and enriched genomic fraction from blood in over 700 subjects. Hum Mol preparation, review, or approval of the manuscript. WW was supported by the Genet. 2014;23:1175–85. Else Kröner-Fresenius Stiftung (2014 A193), the German Research Foundation 11. Christensen BC, Houseman EA, Marsit CJ, Zheng S, Wrensch MR, et al. Aging (WA/1706/8-1), and the Interdisciplinary Center for Clinical Research (IZKF) and environmental exposures alter tissue-specific DNA methylation within the Faculty of Medicine at the RWTH Aachen University (O1-1). dependent upon CpG island context. PLoS Genet. 2009;5:e1000602. 12. Florath I, Butterbach K, Muller H, Bewerunge-Hudler M, Brenner H. Cross- Availability of data and materials sectional and longitudinal changes in DNA methylation with age: an Data protection standards, which were part of the informed consent epigenome-wide analysis revealing over 60 novel age-associated CpG sites. procedure of the ESTHER study, preclude that data can be deposited in Hum Mol Genet. 2014;23:1186–201. Zhang et al. Clinical Epigenetics (2017) 9:9 Page 6 of 6

13. Lin Q, Wagner W. Epigenetic aging signatures are coherently modified in cancer. PLoS Genet. 2015;11:e1005334. 14. Garagnani P, Bacalini MG, Pirazzini C, Gori D, Giuliani C, et al. Methylation of ELOVL2 gene as a new epigenetic marker of age. Aging Cell. 2012;11:1132–4. 15. Horvath S. DNA methylation age of human tissues and cell types. Genome Biol. 2013;14:R115. 16. Hannum G, Guinney J, Zhao L, Zhang L, Hughes G, et al. Genome-wide methylation profiles reveal quantitative views of human aging rates. Mol Cell. 2013;49:459–367. 17. Weidner CI, Lin Q, Koch CM, Eisele L, Beier F, et al. Aging of blood can be tracked by DNA methylation changes at just three CpG sites. Genome Biol. 2014;15:R24. 18. Lin Q, Weidner CI, Costa IG, Marioni RE, Ferreira MR, Deary IJ, Wagner W. DNA methylation levels at individual age-associated CpG sites can be indicative for life expectancy. Aging (Albany NY). 2016;8:394–401. 19. Marioni RE, Shah S, McRae AF, Chen BH, Colicino E, et al. DNA methylation age of blood predicts all-cause mortality in later life. Genome Biol. 2015;16:25. 20. Perna L, Zhang Y, Mons U, Holleczek B, Saum KU, Brenner H. Epigenetic age acceleration predicts cancer, cardiovascular, and all-cause mortality in a German case cohort. Clin Epigenetics. 2016;8:64. 21. Christiansen L, Lenart A, Tan Q, Vaupel JW, Aviv A, McGue M, Christensen K. DNA methylation age is associated with mortality in a longitudinal Danish twin study. Aging Cell. 2016;15:5. 22. Horvath S, Pirazzini C, Bacalini MG, Gentilini D, Di Blasio AM, et al. Decreased epigenetic age of PBMCs from Italian semi-supercentenarians and their offspring. Aging (Albany NY). 2015;7:1159–70. 23. Horvath S, Erhart W, Brosch M, Ammerpohl O, Von SW, et al. Obesity accelerates epigenetic aging of human liver. Proc Natl Acad Sci U S A. 2014; 111:15538–43. 24. Marioni RE, Shah S, McRae AF, Ritchie SJ, Muniz-Terrera G, et al. The epigenetic clock is correlated with physical and cognitive fitness in the Lothian Birth Cohort 1936. Int J Epidemiol. 2015;44:1388–96. 25. Deary IJ, Gow AJ, Pattie A, Starr JM. Cohort profile: the Lothian Birth Cohorts of 1921 and 1936. Int J Epidemiol. 2012;41:1576–84. 26. Tomaszewski M, Debiec R, Braund PS, Nelson CP, Hardwick R, et al. Genetic architecture of ambulatory blood pressure in the general population: insights from cardiovascular gene-centric array. Hypertension. 2010;56:1069–76. 27. Levy D, Ehret GB, Rice K, Verwoert GC, Launer LJ, et al. Genome-wide association study of blood pressure and hypertension. Nat Genet. 2009;41:677–87. 28. Newton-Cheh C, Johnson T, Gateva V, Tobin MD, Bochud M, et al. Genome- wide association study identifies eight loci associated with blood pressure. Nat Genet. 2009;41:666–76. 29. Belsky DW, Caspi A, Houts R, Cohen HJ, Corcoran DL, et al. Quantification of biological aging in young adults. Proc Natl Acad Sci U S A. 2015;112:E4104–4110. 30. BLUEPRINT consortium. Quantitative comparison of DNA methylation assays for biomarker development and clinical applications. Nat Biotechnol. 2016;34:726–37.

Submit your next manuscript to BioMed Central and we will help you at every step:

• We accept pre-submission inquiries • Our selector tool helps you to find the most relevant journal • We provide round the clock customer support • Convenient online submission • Thorough peer review • Inclusion in PubMed and all major indexing services • Maximum visibility for your research

Submit your manuscript at www.biomedcentral.com/submit